TY - GEN
T1 - Discriminative training of GMM via log-likelihood ratio for abnormal acoustic event classification in vehicular environment
AU - Kim, Kwangyoun
AU - Ko, Hanseok
PY - 2011
Y1 - 2011
N2 - In this paper, a discriminative training technique based on Gaussian Mixture Model (GMM) is proposed for detection and classification of abnormal acoustic events in indoor environment. In particular, we consider small indoor space such as vehicular scenes and develop a two-step procedure in which statistical mapping of acoustic features is followed by abnormal event detection. In the first step, Mel-Frequency Cepstral Coefficients (MFCC) feature set is used to construct a Gaussian Mixture Model (GMM) for acoustic event mapping and log-likelihood ratio is used for confidence measure to correct misrecognition over vocal/nonvocal regions. In the 2nd step, an abnormal event is determined using maximum likelihood estimation approach wherein the ratio of abnormal events to cumulative events during an analysis window is compared to a threshold. For performance evaluation, we employ a statistically meaningful database of normal and abnormal acoustic events in actual indoor scenes of two representative scenarios. Subsequent experiments demonstrate a performance of 91% correct detection rate for abnormal context and 2.5% of error detection rate, which indicates it promising for real world vehicular acoustic surveillance applications.
AB - In this paper, a discriminative training technique based on Gaussian Mixture Model (GMM) is proposed for detection and classification of abnormal acoustic events in indoor environment. In particular, we consider small indoor space such as vehicular scenes and develop a two-step procedure in which statistical mapping of acoustic features is followed by abnormal event detection. In the first step, Mel-Frequency Cepstral Coefficients (MFCC) feature set is used to construct a Gaussian Mixture Model (GMM) for acoustic event mapping and log-likelihood ratio is used for confidence measure to correct misrecognition over vocal/nonvocal regions. In the 2nd step, an abnormal event is determined using maximum likelihood estimation approach wherein the ratio of abnormal events to cumulative events during an analysis window is compared to a threshold. For performance evaluation, we employ a statistically meaningful database of normal and abnormal acoustic events in actual indoor scenes of two representative scenarios. Subsequent experiments demonstrate a performance of 91% correct detection rate for abnormal context and 2.5% of error detection rate, which indicates it promising for real world vehicular acoustic surveillance applications.
KW - Acoustic-based surveillance
KW - Context awareness
KW - GMM
KW - MFCC
UR - http://www.scopus.com/inward/record.url?scp=80051657042&partnerID=8YFLogxK
U2 - 10.1109/CNSI.2011.39
DO - 10.1109/CNSI.2011.39
M3 - Conference contribution
AN - SCOPUS:80051657042
SN - 9780769544175
T3 - Proceedings - 1st ACIS/JNU International Conference on Computers, Networks, Systems, and Industrial Engineering, CNSI 2011
SP - 348
EP - 352
BT - Proceedings - 1st ACIS/JNU International Conference on Computers, Networks, Systems, and Industrial Engineering, CNSI 2011
T2 - 1st ACIS/JNU International Conference on Computers, Networks, Systems, and Industrial Engineering, CNSI 2011
Y2 - 23 May 2011 through 25 May 2011
ER -